Multiply robust causal inference with double-negative control adjustment for categorical unmeasured confounding

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Unmeasured confounding poses a significant challenge to establishing causality in observational studies.
• Negative controls, traditionally used for bias detection, are increasingly recognized for causal inference.
• Miao and colleagues previously outlined conditions for using negative controls to identify the average treatment effect (ATE).

Purpose of the Study:

• To establish non-parametric identification of the ATE under weaker conditions using negative control variables and categorical unmeasured confounding.
• To develop a semiparametric framework for ATE inference incorporating numerous measured covariates.
• To propose robust and efficient estimators for ATE estimation.

Main Methods:

• Developed a non-parametric identification strategy for ATE with categorical unmeasured confounding.
• Constructed a semiparametric framework for ATE inference using measured covariates.
• Derived semiparametric efficiency bounds and proposed multiply robust, locally efficient estimators.

Main Results:

• Achieved non-parametric identification of ATE under relaxed conditions.
• Provided a general semiparametric framework for robust ATE estimation.
• Demonstrated the performance of proposed methods through simulations and a vaccine safety study.

Conclusions:

• The proposed methods offer improved causal inference in observational studies with unmeasured confounding.
• The semiparametric framework and estimators provide robust tools for estimating ATE.
• The application to vaccine safety surveillance highlights the practical utility of these techniques.